A Performance Comparison of Hypervisors
22 Pages
English

A Performance Comparison of Hypervisors

-

Downloading requires you to have access to the YouScribe library
Learn all about the services we offer

Description

P E R F O R M A N C E S T U D Y
A Performance Comparison of Hypervisors
VMware A Performance Comparison of Hypervisors
Contents
Introduction....................................................................................................................1
Virtualization Approaches ...........................................................................................1
Enterprise Virtualization Infrastructure ....................................................................2
Test Methodology and Configuration.......................................................................4
Guest Operating System...............................................................................................................................................................4
Test Workloads....................................................................................................................................................................................4
Hardware configuration.......................................................5
Software Configuration..................................................................................................................................................................5
Virtual Machine Configuration ..................................................................................................................................................5
Test Results ............................................ ...

Subjects

Informations

Published by
Reads 47
Language English
A Per
P
E
R
F
O
formance Comparison of Hyper
R
M
A
N
C
E
visors
S
T
U
D
Y
VMware
 A Performance Comparison of Hypervisors
Contents Introduction................................................................................................................... .1  Virtualization Approaches ........................................................................................... 1  Enterprise Virtualization Infrastructure .................................................................... 2  Test Methodology and Configuration ....................................................................... 4  Guest Operating System ......................................................................................................... ...................................................... 4  Test Workloads ................................................................................................................. ................................................................... 4  Hardware configuration ......................................................................................................... ....................................................... 5  Software Configuration......................................................................................................... ......................................................... 5  Virtual Machine Configuration .................................................................................................. ................................................ 5  Test Results.....................................................................................................................6  SPECcpu2000 Integer ............................................................................................................................... ..........................6  Passmark ....................................................................................................................... .......................................................................... 8  Compile Workloads .............................................................................................................. ......................................................... 10  Netperf ........................................................................................................................ .......................................................................... 11  SPECjbb2005 .................................................................................................................... ................................................................. 13  Discussion....................................................................................................................14  Single Virtual-CPU Tests....................................................................................................... ....................................................... 14  Virtual SMP Tests .............................................................................................................. ............................................................... 14  Qualitative Comparison ............................................................................................ 15  Future Work ................................................................................................................. 16  Conclusion...................................................................................................................16  References.................................................................................................................... 17  Appendix A: Test Configurati on .............................................................................. 18  Hardware Configuration ......................................................................................................... ................................................... 18  Client (for netperf tests)...............................................................................................................................................................................18  Native and Guest Operating System Configuration......................................................................................................................18  Native:....................................................................................................................................................................................................................18  VMware ESX Server 3.0.1..............................................................................................................................................................................18  Xen 3.0.3-0.................................................................................................................... ....................................................................................... 18  Hypervisor Configurations..........................................................................................................................................................................19  
  
Contents
 
 
  
i  
VMware
 
 A Performance Comparison of Hypervisors
Introduction   Virtualization has rapidly attained ma instream status in enterprise IT by delivering transformative cost savings as well as increased operational efficienc y, flexibility and IT ser vice levels. Intel and AMD have independently developed virtualization extensions to the x86 architecture called hardware virtualization. This and other recent hardware advances such as multicore processors are further fueling the adoption of virtualization. While a full vir tual ser vice-oriented infrastructure is composed of a wide ar ray of technologies that provide resource aggregation, management, availability and mobility 1 , the foundational core of virtual infrastructure is the hypervisor. This paper provides a quantitative and qualitative comparison of two virtualization hypervisors available for the x86 architecture — VMware ESX Server 3.0.1 and open-source Xen 3.0.3 — to validate their readiness for enterprise datacenters. A series of performance experiments was conducted on the latest shipping versions (at the time of this study in November 2006) for both hypervisors using Microsoft Windows as the guest operating system. This white paper discusses the results of these experiments. The discussion in this white paper should help both IT decision makers and end users to choose the right virtualization hypervisor for their datacenters. The experimental results show that VMware ESX Server delivers the superior, production-ready per formance and scalability needed to implement an efficient and responsive datacenter. Furthermore, while we had no problems exercising enterprise virtualization capabilities such as Virtual SMP and virtual machine scalability using the VMware ESX Ser ver hypervisor, we were not successful in running similar tests with the Xen 3.0.3 hypervisor due to product failures. Virtualization Approaches The x86 architecture is the most popular computer architecture in enterprise datacenters today, hence virtual infrastructure for the x86 architecture has tremendous ben efits. The two leading software virtualization approaches to date have been full virtualization and paravirtualization. AMD and Intel have recently introduced new processor instruc tions to assist virtualization software. The full virtualization approach allows datacenters to run an unmodified guest operating system, thus maintaining the existing investments in operating systems and applications and providing a nondisruptive migration to virtualized environments. VMware uses a combination of direct execution and binary translation techniques [1] to achieve full virtualization of an x86 system The paravirtualization approach modifies the guest operating system to eliminate the need for binary translation. Therefore it offers potential performance advantages for certain workl oads but requires using specially modified operating system kernels [2]. The Xen open source project was designed initially to support paravirtualized operatin g systems. While it is possible to modify open source operating systems, such as Linux and OpenBSD, it is not possible to modify “closed” source operating systems such as Microsoft Windows . It is also not practical to modify older versions of open source operating systems that are already in use. As it turns out, Microsoft Windows is the most widely deployed operating system in enterprise datacenters. For such unmodified guest operating systems, a virtualization hypervisor must either adopt the full virtualization approach or rely on hardware virtualization in the processor architecture.                                                  1 Resource aggregation refers to the capability to pool, share, and throttle memory, processing power, network, and storage across server instances. Mobility refers to the capability to perform live migrations of running virtual m achines from one physical server to another in response to availability requirements.
Introduction
 1
VMware
 A Performance Comparison of Hypervisors
The hardware virtualization support enabled by AMD-V and Intel VT technologies introduces virtualization in the x86 processo r architecture itself. While first-generation h ardware assist support includes CPU virtualization only, later generations are expected to include memory and I/O virtualization as well. The emergence of virtualization hardware assist reduces the need to paravirtualize guest operating systems. In fact, Xe n vendors such as Virtual Iron have announced that they are supporting only full virtualization using AMD-V and Intel VT processors and are not supporting paravirtualization [9]. While an architectural comparison between these approaches is of interest to those trying to predict the long-term direction of virtualization technology, the advantage of any one approach for any single element of virtualization overhead may be outweighed by a variety of other datacenter requirements outlined in the “Datacenter Requirements” section of this paper. It is a combination of all three approaches that will ultimately help architect a successful virtual datacenter. Enterprise Virtualization Infrastructure Enterprise datacenters typically start by implementing virtualization as the basis for server consolidation and containment 2 . Over time, IT staff tend to branch out in their use of virtualization, to the point where it becomes a standard part of the production datacenter infrastructure. While this standardization on vir tual infrastructure provides tremendous value in improved resource utilization, superior manageability and flexibility, and increased application availability, these benefits cannot be achieved through the hypervisor alone. Enterprise virtualization is a broad IT initiative, of which basic server partitioning is just one facet.  
                                                 2 Consolidation is the process and result of shrinking the overall server footprint in a datacenter to a smaller number of virtualized servers. Containment is the process and result of containing the further proliferation of physical server s, beginning at a particular time, through virtualization.
Enterprise Virtualization Infrastructure
 2
VMware
Infrastructure Optimization
Business Continuity
Core Management Automation, Tools
System Infrastructure Services
Infrastructure Virtualization
 A Performance Comparison of Hypervisors
SW Lifecycle Automation
Virtual Clients and Desktops
Single Node Hypervisor  Figure 1 — Enterprise Virtualization Infrastructure  As illustrated in Figure 1, enterprise virtualization infrastructure consists of the following components:  Single-node hypervisor  to enable server partitioning capability.  Infrastructure virtualization that virtualizes and aggregates industry standard servers and their attached network and storage into unified resource pools.  A set of virtualization-based distributed system infrastructure services such as resource management to dynamically and intelligently optimiz e the available resources among virtual machines. High availability for better service levels, data protection for reliable and cost effective disaster recov ery, and security and integrity to better protect e xisting infrastructure investments from typical datacenter vulnerabilities. A suite of management automation technologies and tools that provide virtualization-specific capabilities such as comprehensive system resource monitoring (of metrics such as CPU activity, disk access, memory utilization, and network bandwidth), automated provisioning, cloning, and workload migration support.  A set of end- to-end solutions such as infrastructure optimization, business continuity, software lifecycle automation, and Virtua l Desktop Infrastructure comple te the virtual infrastructure.   
Enterprise Virtualization Infrastructure
 3
VMware
 A Performance Comparison of Hypervisors
Together, all these components in the enterprise virtualization infrastructure are required to successfully implement virtualization inside datacenters. The foundational element of the virtual infrastructure, however, is the hypervisor. The next two sections provide a comparison of the operational characteristics of the VMware ESX Server 3.0.1 and Xen 3.0.3 hypervisors, with a specific focus on their respective per formance characteristics. Test Methodology and Configuration To conduct a quantitative comparison of hypervisors, a few key decisions had to be made — th e choice of the guest operating system and the workloads to use for the evaluation. Guest Operating System Microsoft Windows Server 2003 was selec ted as the guest operating system for these tests for several reasons. First, Microsoft Windows operating systems are the most widely deployed operating systems on x86 platforms. Second, typical enterprise customers run standard, off-the-shelf operating systems and sof tware in their virtual machines to maintain compatibility and compliance with their suppor t agreements. To run such unmodified guest operating systems with the Xen 3.0.3 hypervisor, one needs to use the latest generation of hardware that suppor ts virtualization in the x86 processors. The Linux community has adopted the “para-virt ops” API approach based on VMware’s proposed Virtual Machine Interface ( VMI), a completely open hypervisor interface. The “para-virt ops” API together with back-end VMI support is scheduled to be included in the Linux kernel version 2.6.21. VMware has already announced plans to support these paravirtualized guest operating systems and we will revisit these tests for both unmodified and paravirtualized Linux guest operating systems at that time. Test Workloads A typical enterprise datacenter runs a mix of CPU-, memor y-, and I/O -intensive applications. Hence the test workloads chosen for these experiments comprise several well-known standard benchmark tests, as listed below: The integer component of the SPECcpu2000 benchmark suite, available from SPEC® (Standard Performance Evaluation Corporation), was chosen to represent CPU-intensive applications [6]. Passmark, a synthetic suite of benchmarks intended to isolate various aspects of workstation performance, was selected to represent desktop-oriented workloads [4] . Netperf was used to simulate the network usage in a datacenter [3]. The SPECjbb2005 benchmark suite from SPEC was used to represent the Java applications typically used in the datacenters [7]. A compile workload — build SPECcpu2000 INT pack age — was also added to capture typical IT development and test usag e in datacenters. The objective of these experiments was to test the performance and scalability of the two virtualization hypervisors. The tests were per formed using a configuration with a single virtual CPU. We attempted to repeat the single-virtual-CPU tests using virtual SMP configurations (for example, two virtual CPUs and four virtual CPUs), as well as to run scalability tests using multiple virtual machines. However, it was not possible to run the Xen 3.0.3 hypervisor either in multiple
Test Methodology and Configuration
 4
VMware
 A Performance Comparison of Hypervisors
virtual CPU configurations or using multiple virtual machines. More details have been provided later in the paper. It must be noted here that any experimental test setup based solely upon resource-intensive workloads driving a physical system into and past saturation is not a likely customer scenario. In most production deployments, IT managers conduct detailed capacity planning and sizing exercises and the average utilization of the servers is kept within reasonable limits to allow for usage spikes and future capacity growth. The benchmar k test suites are used in these experiments only to illustrate per formance and scalability of the two virtualization hypervisors. Hardware configuration The system used to run all benchmark tests was an IBM X3500 ser ver with two VT-enabled dual-core 3GHz Intel Woodcrest CPUs (total four cores). Although the test system had 5GB of RAM installed, it was booted with only 1GB of RAM for native tests. Additionally, the test system was configured with a dualpor t 1Gbps Ethernet adapter and t wo 146GB SAS disk dr ives. For native operating system tests, all data was captured using Windows Ser ver 2003 Enter prise Edition R2 32-bit. Only the Netper f tests needed a client, which used either one or two Microsoft Windows 2000 clients. The client used was a Dell 1600SC ser ver configured with two Pentium 4 2.4GH z processors and one 1Gbps network adapter card. All tests were controlled from within the virtual machine itself. As shown in Figure 2, both Netperf clients communicated with a single virtual machine.
 Figure 2 — Configuration for two-client Netper f test Software Configuration All the experiments described in this paper were run using ESX S erver 3.0.1 GA release and Xen 3.0.3-0 release. Both were the latest shipping releases for the two virtualization hypervisors at the time of this testing in November 2006. We downloaded the Xen 3.0.3 version from University of Cambridge Computer Laboratory [8]. Virtual Machine Configuration Each virtual machine was configured with one virtual CPU and 1GB of memor y unless specifically noted. For the SPECjbb2005 tests, each virtual machine was configured with 1.6GB of memor y and two or four virtual CPUs based on the test run. The Windows Server 2003 EE R2 32-bit operati ng
Test Methodology and Configuration
 5
VMware
 A Performance Comparison of Hypervisors
system was installed inside the virtual machine. The 32-bit version was chosen because it is still the most deployed operating system. We plan to run similar tests using the 64-bit version in the future. No attempt was made to optimize the benchmark test results in any way. Default tools and settings were used in all cases. For SPECjbb2005 tests, BEA Systems’ JRockit 5.0 R26.4.0-63 Java virtual machine ( JVM) was used. The java virtual machine options for SPECjbb2005 tes ts were set to -Xms960m -Xmx960m -Xgc:parallel -XXaggressive:opt -XXcompactratio8 -XXminblocksize16k . For each test run, only the single active virtual machine was powered on, since idling virtual machines continue to consume a small amount of resources and can sk ew results. Test Results This section provides detailed results for each of the experiments. All results, unless specifically noted, have been normalized to native per formance on a throughput basis to make it easier to illustrate the slowdown resulting from virtualization. Higher numbers indicate better per formance, unless indicated otherwise. For an enterprise datacenter, better per formance implies significant benefits in several ways: Better performance across different wor kloads implies that more application types can be successfully deployed in production in virtual environments. Near-native performance al so indicates that more virtual machines can be deployed on a single physical server, resulting in higher consolidation ratios. This can help even if an enterprise plans to standardize on virtual inf rastructure for server consol idation alone. Finally, better performance can have a measurable impact on many of the costs, such as hardware, software, administration, support, system downtime, and user productivity, which influence t he total cost of ownership (TCO). Hence a hypervisor ’s performance can contribute considerably towards easing the initial cultural shift to adopting virtualization inside datacenters as well as transparently migrating end users to virtualized environments. However, as stated earlier, a hypervisor is just one component required for successfully implementing enterprise virtualization infrastructure.
Test Results
 
 6
VMware
 A Performance Comparison of Hypervisors
SPECcpu2000 Integer   This benchmar k comprises mostly user-level computation, hence we expect both virtualization hypervisors to score close to native. The results, as shown in Figure 3, show a slowdown over native ranging from 0–6 percent for the VMware ESX Ser ver and from 1–12 percent slowdown for the Xen 3.0.3 hypervisor. Overall, Xen 3.0.3 shows twice the overhead of VMware ESX Server, an average slowdown of 6 percent compared to 3 percent.  
1.00 0.90 0.80 0.70 0.60 0.50 0.40 0.30 0.20 0.10 0.00
Native ESX301 Xen3030 Figure 3 — SPECcpu INT 2000 results compared to native (higher values are better)
Test Results
 
 
 
7
VMware
 A Performance Comparison of Hypervisors
Passmark Figure 4 shows the results obtained for CPU tests in the Passmark benchmar k suite. The following CPUmark subtests were run during these experiments: IntMath, FPMath, MMX, SSE/3DNow, Compression, Encryption, ImageRotate, and StringSort. These tests comprise mostly user-level computation, hence we expect both virtualization hypervisors to score close to native. The results show a slowdown over native ranging from 4–18 percent for VMware ESX Server and from 6–-41 percent overhead compared to native for the Xen 3.0.3 hypervisor. Overall, Xen 3.0.3-0 shows almost twice the overhead, an average slowdown of 17 percent compared to 9 percent for VMware ESX Server.  
1.00 0.90 0.80 0.70 0.60 0.50 0.40 0.30 0.20 0.10 0.00
Native ESX301 Xen3030  Figure 4 — Passmark – CPU results compared to native (higher values are better)  Both SPECcpu2000 I nteger and Passmark – CPU tests demonstrate that VMware ESX Ser ver can handle CPU intensive applications — such as database servers, application ser vers, file ser vers, terminal servers, and mail servers — in a typical enterprise datacenter more efficiently than the Xen hypervisor. Figure 5 shows the results for Memor y tests in the Passmark benchmar k. The Memor ymark subtests included: AllocateSmallBlock, ReadCached, ReadUncached, and Write. Both VMware ESX Server and Xen hypervisors demonstrate near nat ive per formance. The VMware ESX Server shows an average 2 percent overhead compared to native, while the Xen results show an average of 3 percent overhead compared to the native performance.  
Test Results
 
 8
VMware
 
 A Performance Comparison of Hypervisors
1.00 0.90 0.80 0.70 0.60 0.50 0.40 0.30 0.20 0.10 0.00 Allocate Small Read Cached Read Uncached Write Memory Mark Block Native ESX301 Xen3030 Figure 5 — Passmark - Memory results compared to native (higher values are better)
Test Results
 
 
 
9